Open cczhu opened 4 years ago
cczhu
commented
4 years ago Sensitivity testing: given variations in daily count totals at a short term count location, how much does CountMatch's AADT prediction vary? This can be tested either by varying STTC data by X%, or by "swapping" data between neighbouring STTCs.
The results of this experiment can also point to where we are in greatest need for PTCs, helping to inform the 2020 count program.
cczhu
commented
4 years ago Deviation from model assumptions: since we calculate an MSE or COV to match an STTC with a nearby PTC, the goodness of match is encoded by the value of the error metric. Both metrics are technically [0, infty), but MSE or COV of >~ 1 is probably a terrible fit. We can use this mismatch as goodness of fit metric between the model and the data. It wouldn't necessarily indicate that there's something wrong with either, just that the data and model assumptions don't align for whatever reason.
Traffic Prophet currently checks CountMatch's predictive accuracy by creating a set of fake STTC out of our PTC. TEPs checks KCOUNT and LSVR accuracy by using CountMatch's outputs as ground truth. While these methods are useful for spotting eg. poor fits in KCOUNT or CountMatch, they have extremely limited ability to determine the predictive error on roads where we don't already have empirical AADTs.
We need to brainstorm ways of estimating the predictive accuracy of Traffic Prophet, ideally in ways that minimize the data bias we currently have to expressways.